0.2.0 release

FastJev v0.2.0 — what’s new since v0.1.1:

• Cross-state batching: 57% higher throughput on our mixed-length RTX 5090 / Qwen3.5-4B benchmark
• Torch/MPS and EXL3 backends for Apple GPUs and larger quantized models
• Workload-specific temperature calibration and optional llama.cpp prefix reuse
• Chinese UI for the WebGPU demo
• this-that model comparisons, per-decision results, and reproducible benchmarks

Full Changelog: v0.1.1...v0.2.0

v0.1.1

What's Changed

• docs(readme): remove independent iteration claim by @chengyongru in #11
• docs(readme): tighten project rationale by @chengyongru in #12
• docs(readme): add FastJev wordmark by @chengyongru in #14
• feat: add llama.cpp GGUF backend by @chengyongru in #13
• Focus README on Jev self deployment and measured results by @chengyongru in #15
• Prepare fastjev 0.1.1 for release by @chengyongru in #16

Full Changelog: v0.1.0...v0.1.1

fastjev v0.1.0

Release notes

fastjev v0.1.0 is the first release of an independently maintained fork of TheoLeeCJ/SemIf.

This fork prioritizes an SDK-first integration path, independent iteration, and faster inference for self-hosted semantic decisions.

Highlights

• Typed Python SDK for Boolean, Choice, and Score decisions
• Backend-neutral ScoringBackend protocol
• Built-in Torch, MLX, and optional vLLM backends
• Batched vLLM inference without coupling the public SDK to one runtime
• System One-compatible HTTP adapter
• Reproducible quality, perturbation, and performance evidence
• English and Simplified Chinese documentation

Install

Torch/CUDA:

pip install "fastjev[torch]"

vLLM:

pip install "fastjev[vllm]"

HTTP API with Torch:

pip install "fastjev[api,torch]"

Apple Silicon / MLX currently requires the validated MLX-LM source revision because the latest published MLX-LM 0.31.3 release does not contain the required Q/K
normalization fix:

pip install "fastjev[mlx]" \
  "mlx-lm @ git+https://github.com/ml-explore/mlx-lm.git@a63e24c389382619eb6d9af656e3b46024be217a"

On the same RTX 5090 host, model revision, and three-decision safeguard batch:

• Torch: 18.52 decisions/s, 161.97 ms median batch latency
• vLLM: 36.26 decisions/s, 82.73 ms median batch latency
• vLLM delivered 1.96× throughput and 48.9% lower warm-batch latency

Quality comparison:

• Authored set: 144/144 argmax agreement
• Perturbation set: 106/108 argmax agreement

These measurements are scoped to the documented hardware, model, inputs, and timing method.

Validation

• 67 tests passed; one platform-gated MLX suite skipped on non-Apple hardware
• Wheel and source distribution passed twine check
• Distribution metadata passed the PyPI direct-dependency guard
• Raw evidence checksums and 69 published claims were verified
• fastjev==0.1.0 was published to TestPyPI and installed successfully in a fresh environment

Scores are conditional distributions over declared options, not calibrated confidence estimates.